| tags: | |
| - time-series | |
| - kinematics | |
| - biomechanics | |
| - contrastive-learning | |
| - gait-analysis | |
| - transformer | |
| license: apache-2.0 | |
| library_name: pytorch | |
| pipeline_tag: time-series-classification | |
| # PoseT5: Biomechanical Gait Kinematics & Keystroke Predictor (CAL Trained) | |
| PoseT5 is a lightweight (~430K parameter) time-series transformer architecture trained with Supervised Contrastive Alignment Loss (CAL) to predict inverse keystrokes from 14-channel invariant biomechanical joint angles and velocities. | |
| ## Organization | |
| Trained and hosted under **[Learningbase](https://huggingface.co/Learningbase)**. | |
| ## Performance Metrics | |
| - **Validation Macro F1**: 0.7378 | |
| - **Hamming Loss**: 0.0317 | |
| | Keystroke | Action | F1 Score | Precision | Recall | | |
| |---|---|---|---|---| | |
| | `W` | Forward Walk/Run | 0.970 | 0.965 | 0.976 | | |
| | `A` | Yaw Turn Left | 0.881 | 0.839 | 0.929 | | |
| | `S` | Backward Move | 1.000 | 1.000 | 1.000 | | |
| | `D` | Yaw Turn Right | 0.837 | 0.947 | 0.750 | | |
| | `Space` | Jump Dynamics | 0.000 | 0.000 | 0.000 | | |
| ## Architecture & Loss | |
| - **Input Channels**: 14 invariant biomechanical joint angles and angular velocities. | |
| - **Instance Normalization**: Reversible Instance Normalization (RevIN). | |
| - **Temporal Patching**: Slices 32 frames into 4 patches (P=8) projected into d_model = 128. | |
| - **Contrastive Alignment Loss (CAL)**: InfoNCE loss aligning character gait embeddings across different rigs and morphological dimensions. | |
| - **Backbone**: 2-layer Multi-Head Self-Attention Transformer Encoder (~430K parameters). | |
| - **Latency**: Sub-millisecond CPU inference (< 2ms). | |